Because a simulator never matches reality exactly, sim-to-real pipelines close the gap with domain randomization: physical parameters such as friction, actuator dynamics, command delay, and battery voltage are varied across thousands of parallel simulated environments so the learned policy stays robust when it meets real hardware. Pollen Robotics' Microduck is a consumer-priced example — PPO policies trained in MuJoCo, exported to ONNX, and run on the robot at 50 Hz.